Vector Embeddings: Meaning, Similarity, and Model Compatibility

An embedding represents an input as a list of numbers that a model has learned to place in a useful geometric space. Similarity between those vectors can help retrieve related material, but the numbers are not a universal encoding of meaning.

The model, preprocessing, task, and distance measure all matter. Two vectors with the same number of dimensions are not necessarily comparable if different models produced them.

Think of the space as a shared coordinate system

A query and its candidate documents need compatible representations. If a provider requires different query and document prefixes, use them consistently. If you switch models, assume you need a deliberate migration rather than mixing old and new vectors in one index.

The embedding-model evaluation guide explains how to record model revisions and preprocessing. Keep those details with the stored records, so an unexpected retrieval change can be traced to a representation change.

Similarity answers a limited question

A close vector suggests related content under that representation. It does not establish that the passage is current, permitted, true, or sufficient to answer the question. An old cancellation policy can be semantically closer than the current policy's differently worded exception.

Apply access and version rules before presenting evidence to the answer model. Use exact matching or structured lookup when the task depends on a particular identifier or authoritative record.

Choose the metric with the model and index

Cosine similarity compares direction. Dot product also reflects vector magnitude unless the vectors are normalized. Euclidean distance measures separation in the coordinate space. Follow the model and index requirements rather than choosing a metric because its values look intuitive.

A score is not automatically a probability that the answer is correct. A threshold calibrated for one model and corpus may behave differently after a model, preprocessing, or content change.

Test a model change as a retrieval change

Build a parallel index for a representative sample. Run exact names, paraphrases, ambiguous questions, and no-answer cases against both versions. Inspect ranked source IDs as well as final prose, and keep the old representation available while comparing.

Avoid attributing a ranking change to the model if chunking and metadata filters changed at the same time. The vector-search guide connects representations to indexing and retrieval.

For repository search, the code-embedding dimension experiment shows how to compare smaller vectors without assuming recall is unchanged.

If you want to compare that work with a managed context layer, start a Supermemory pilot using the same source set and questions. Compare supported answers and operating effort, not embedding dimensions alone.